This project aims to do the first step towards building a simple speech detector using Speech Commands Dataset released by TensorFlow.

The dataset used in this project is the Speech Commands Dataset by TensorFlow. The dataset contains 65,000 one-second long utterances of 30 short words and a separate folder with backgound noise audio clips. These clips were broken into one-second chunks and separated into a folder named 'silence', for classification.

Mel-Frequency Cepstral Coefficients (MFCC), Chroma and Contrast features were extracted and used as features.

The features were normalized and PCA was performed on them to reduce dimension.

A 3 layer deep neural network was implemented in TensorFlow to do the classification. After experimenting with a few combinations I chose the hidden layer units as 700, 700 and 100. The output of each hidden layer passes through a RELU activation. Score - 0.63

A Random Forest model was fit with the default parameters. Score - 0.64

kNN was used to fit the data with the default parameters. Score - 0.65

The result obtained is not the best but is in the top 600 results obtained by other competitors in Kaggle.

I attribute this to the fact that the audio clips with background noise need more processing before their features should be extracted. Also, the words that are classified as 'unknown' are not similar to each other but are similar to the other classes of words.